Seedream 5.5 Beats GPT Image 2.5 Sunburst on Text — On ByteDance's Own Scoreboard

Two numbers carry the text-rendering story in Seedream 5.5, and neither of them is a benchmark score in the ordinary sense. ByteDance says severe text corruption lands at 0.4% and minor corruption at 1.7% across Chinese and English, including dense small type, and that its own evaluations put Seedream 5.5 ahead of both GPT Image 2.5 Sunburst and Nano Banana 2.1 on text structure in text-to-image and image-to-image. That is a specific, falsifiable claim about the thing that breaks most often in commercial image work. It is also, as of 10 November 2026, one day after the model was announced, entirely unverified by anyone outside ByteDance — and the two models it names sit at the top of an independent leaderboard that has no row for it at all.

What the 0.4% is actually measuring

Text corruption rates are the rare image metric with a definition you can argue about productively. A rendered headline is either legible or it is not. A one-character error in a brand name is a reshoot; a swapped character in a price is worse than a reshoot. So when a vendor publishes "severe 0.4%, minor 1.7%", the interesting questions are which character sets were tested, at what size, and who graded "minor".

ByteDance's announcement answers the first two: Chinese and English, dense small text included, evaluated across six use cases on its own MagicArena setup. It does not answer the third in public, and it does not publish per-model breakdowns that would let an outsider recompute the comparison against GPT Image 2.5 Sunburst or Nano Banana 2.1. That is not unusual — almost nobody publishes grading rubrics — but it does mean the number is a claim about ByteDance's process as much as about the model.

What makes the claim worth taking seriously anyway is that it is narrow. "Better text structure" is a much smaller promise than "better images", and narrow promises are the ones vendors usually get right, because they are the ones their own customers complain about. The commercial logic is easy to follow: ByteDance's consumer surfaces — Doubao, Doubao Work, Dreamina, the BytePlus AI Playground, Dola — are full of posters, product cards and social assets, and every one of those is a text-rendering test.

The rivals it names are on the board. It is not.

Here is the awkward part of the comparison. On Artificial Analysis's image editing leaderboard, version 2.1, GPT Image 2.5 Sunburst sits third at 1,199 Elo and Nano Banana 2.1 sits seventh at 1,152, both with confidence intervals around ±12 points. On the text-to-image board, GPT Image 2.5 Sunburst is third at 1,227. Those are blind human-preference measurements with thousands of votes behind them and no vendor involvement.

Seedream 5.5 has none of that. Neither board carries a row for it. The nearest thing to an independent anchor is its predecessor: Seedream 5.0 Pro at 1,129 Elo on editing (eleventh) and 1,101 on text-to-image (eighteenth). So the comparison ByteDance is inviting — new model versus the two strongest models on text — is a comparison between a vendor's internal arena result and two independently sampled Elo scores. They are not the same kind of measurement, and putting them in one sentence without saying so is how vendor numbers become folklore.

That gap will close or it will not. Artificial Analysis typically adds rows within days to weeks of a model becoming API-accessible, and Seedream 5.5's developer API is still described as opening "this week" on Volcano Ark, with no public model identifier in ModelArk's documentation as of today. Until a row appears, the honest reading of the text claim is: plausible, specific, and untested.

Why text rendering is the claim that matters commercially

Most image-model comparisons are about taste, and taste is a bad basis for a purchasing decision. Text is different. A model that renders a Chinese film title in Traditional characters instead of Simplified is not a model with a style problem; it is a model that cannot be used in a localization pipeline. That failure mode has been documented repeatedly in independent testing of rival models, and it is the single most common reason a design team keeps a model out of production even when the pictures look good.

If ByteDance's numbers hold even loosely — say, corruption rates in the low single digits rather than sub-percent — Seedream 5.5 becomes viable for a category of work that has been awkward for every image model: poster and packaging mockups where the type has to survive, e-commerce cards where the price and the SKU both have to be right, and any bilingual asset where one model has to handle two scripts in one frame.

If they do not hold, the model still has two other things to sell, and both are covered elsewhere in this series: a native alpha channel, and an editing loop ByteDance claims survives a hundred rounds without drifting. Text is the headline; those are the features.

The endurance claim is the one you can test this afternoon

There is a useful asymmetry in this launch. The text numbers require an evaluation harness, a character-level diff and a grading rule. The endurance claim requires none of that. ByteDance says unedited regions are preserved at the pixel level and that the same image survives 100 rounds of editing. You can open Dreamina, take one product photo, and run twenty edits on it while watching a region you never touched. Either the shadow behind the subject is the same shadow in round twenty or it is not.

That is worth doing before you trust any of the published percentages, because the two claims are not independent in practice. A model that holds an image together across a long session is a model whose text also survives being edited around — the type does not get re-rendered from scratch on every turn. A model that drifts will eventually drift through its lettering, and no single-pass corruption rate will tell you that.

Where a routing layer helps when the model is unscored

An unscored model is a procurement problem before it is an engineering problem. You want to try Seedream 5.5 on real briefs without rewriting a pipeline around an endpoint whose model identifier does not exist yet, and you want the option to walk away if the text numbers turn out to be a vendor-arena artefact.

That is the case OrcaRouter is built for, with one honest caveat: we do not carry any ByteDance image model, so this launch is not something you can call through us on day one. What a single endpoint does buy you is the comparison. OrcaRouter passes provider list prices through at 0% markup across 200-plus text and image models, which means the model you benchmark Seedream 5.5 against — GPT Image 2.5 Sunburst, Nano Banana 2.1, Imagen 4.0 — is priced on your invoice at exactly what its vendor charges, and a vendor price change is live on our side the same day rather than at renewal. When the Volcano Ark rate card does land, the number you hold it against is the vendor's own, not a reseller's.

Failover matters for a different reason. A brand-new endpoint in its first week is the least reliable it will ever be: rate limits unannounced, latency unpredicted, refusals undocumented. Routing a percentage of traffic to it with an automatic fallback to a model you have already scored is how you get evidence without betting a production path on a launch-week API.

What would change this verdict

Three things, in order of how much they would move the answer.

First, a row on the editing board. Seedream 5.5 is an editing-first model — the hundred-round claim is about the editing loop — so the editing board is where it should show up strongest, and where a gap against GPT Image 2.5 Sunburst's 1,199 would be most informative. If it lands near its predecessor's 1,129 while holding the loop, it is a workflow tool at a good price. If it lands above 1,180, the text claim becomes much easier to believe and the incumbents' pricing becomes the story.

Second, somebody reproducing the corruption rates. A character-level diff over a few hundred prompts in both scripts is a weekend of work for anyone with an eval harness, and the first person to publish one will settle the question the launch was built on.

Third, the Pro variant ByteDance says will follow the 5.5 series. Buying decisions made against today's model should assume a more expensive, more capable sibling lands on top of it, which is the normal pattern and a reason to keep the model identifier in configuration rather than in code.

Until one of those lands, the fair summary is that ByteDance has made the most specific text claim in the image market this month, attached it to a model that is live in its consumer apps today, and left the verification to everybody else. That is not a criticism so much as a description of how image launches work in 2026. The claim is worth tracking precisely because it is checkable — and the check is cheaper than the argument.